You are on page 1of 18

Credit Risk Mitigation Evidence in Auto Leases: LGD and Residual Value Risk

Hugues Pirotte, Mathias Schmit and Cline Vaessen This version: January 12th, 2004

Abstract This paper is devoted to recovery and residual value risks modeling issues of automotive lease portfolios. First, loss given default distributions are estimated and compared for dierent samples based on risk drivers. Secondly, residual value risk is approached through a re-sampling technique to provide the rst empirical residual value losses in the automotive lease sector. Probability density function of losses and VaR measures are estimated on the basis of a private database comprising a unique set of 4,828 individual automotive lease contracts issued between 1990 and 2001 by a major European nancial institution. Then, a discussion is led in relation to the capital requirements related to residual value risk stemming from the Basel Committees proposed new framework (as in CP3). As the greatest part of recovery risk is diversiable, our conclusion is that a wider recognition of physical collateral under Basel II should allow to better reecting the relatively low-risk prole of automotive lease exposures. JEL classication: G21;G28;G33. Keywords: Credit risk, Collateral, Basle Accord, Capital requirements, Retail Financing, Leasing.

Introduction

Little research has been carried out on the risk of retail lease portfolios although the volume of new business in the lease-nancing sector rose to more than C 199 billion = according to Leaseuropes 2002 estimates, i.e. the penetration rate of equipment lease in comparison with total equipment investments reached 15%. Indeed the studies are driven by the construction of databases comprising proprietary data that allow us to run eective risk models. In fact, retail markets present a number of challenges for credit risk estimations (see Allen, Delong and Saunders [1]) since retail portfolios are illiquid and are not traded in secondary markets; furthermore, borrowers tend to be informationally opaque and borrow infrequently.
This paper has beneted from the data support of a major Belgian nancial leasing company. Research Center E. Bernheim, Solvay Business School, University of Brussels, av F.D. Roosevelt 21, CP 145/01, B-1050 Brussels, Belgium. Corresponding authors email: cvaessen@ulb.ac.be

Credit Risk Mitigation in Auto Leases

Four empirical studies have recently been conducted to assess credit risk in the leasing business. Though based on a relatively small amount of data, De Laurentis and Geranio [9] provide useful empirical and quantitative information, suggesting, in particular, that the European leasing industry benets from high recovery rates in the event of default. Working with a much larger sample of individual defaulted lease contracts issued between 1976 and 2002 by 12 companies in six dierent countries, Schmit and Stuyck [20] extended the investigation to include an analysis of recovery rates relative to the age, term-to-maturity, and default date of each contract. Their study conrmed De Laurentis and Geranios earlier nding that leasing companies incur relatively low losses when a lease defaults. A later study by Schmit [18] estimates the probability density function of losses and VaR measures in a portfolio of vehicle leases issued between 1990 and 2000 by a major European nancial institution. The estimates are carried out on a model based on CreditRisk+TM (Credit Suisse Financial Product [8]). The results of the rst study are conrmed by the second, which applying a non-parametric simulation extends the scope of the investigation. In fact, in addition to the above-mentioned automotive lease contracts, the proprietary database used for the second study includes leases for oce equipment/computers, medical equipment and other kinds of equipment. Sub-portfolio losses for each type of asset are then estimated with a re-sampling or bootstrap technique similar to the one used by Carey [5] to estimate credit losses in private debt portfolios. The results of these two latest studies suggest that the capital requirements prescribed under the current Basel Committees [2] proposed new framework (as in CP3) are excessive for automotive lease business. This paper goes beyond credit risk to focus specically on recovery risk and the role of physical collateral in lease portfolio. Even though it is a common feature, little eort has been empirically devoted to the issue. Indeed, we propose to assess the value of physical collateral on a given automotive lease portfolios representative of its market. Results are provided for several subsamples, based on the nancial remaining duration (which comes out to be particularly useful and trustworthy, as opposed to calendar duration) and dierent states of the economy in order to assess whether recovery risk is diversiable or not. As an extension, we calculate the distribution of losses incurred on automotive lease portfolio associated with residual value risk i.e. loss experience when the realized value of a lease asset at its maturity is less than the option price dened at the beginning of the lease. This is particularly useful in respect to the discussion led in relation to the capital requirements related to residual value risk stemming from CP3. The next section describes the data, proposes some computations of key risk variables while a perspective on the stratication of the data across is presented. This will be of main importance when determining which drivers will be cross-checked against our loss and recovery risk estimations. Section 3 outlines our methodology to estimate recovery risk in addition to explaining the re-sampling technique used for the calculation of loss distribution tails. Section 4 provides empirical results and discuss them against some regulatory implications. Finally, a conclusion is drawn.

Credit Risk Mitigation in Auto Leases

The Data

Lease denition encompasses various types of contracts. In the current research, lease contracts are mainly non-cancellable and lessees are responsible for the selection, acquisition and maintenance of the asset. At maturity, the residual value of the leased asset returns to the lessor but the lessee has usually the right to buy it for a contractual residual value. Our database consists of a unique set of 4,828 individual defaulted auto lease contracts issued between 1990 and 2001 by a major European leasing company with an over 20% share of its national market. All leases are completed contracts. The database contains all the relevant information concerning the leases throughout their life. The available variables fall into two categories: ex-ante and ex-post. Ex-ante variables are the origination date of the contract, the cost of the asset, the contractual maturity of the lease, the periodicity of forecasted payments and the asset description. As regards ex-post variables, we have comprehensive data concerning the date of the default declaration, the total amount already invoiced but not honoured (called total overdue", an annuity including accrued interests and a fraction of the principal amount), the amount of outstanding capital remaining after the last defaulted invoiced amount, the amount recovered and its eective date thanks to the asset sale (free of additional executed guarantees), the charged-o date (date of nal accounting loss recognition), the procedure length (in general, calendar time between charged-o and default dates, the yearly yield of the contract alive (may be used as an opportunity cost). Given the previous denitions, it must be noted that at the default date, the total amount due by the counterpart is the sum of the outstanding capital and the total overdue", such that Total due = Capital Outstanding + Total overdue (1)

Figure 1 presents the relationship between the principal and interest components of annuity payments and these observed variables.

2.1

Descriptive statistics

Table 1 provides a sampling of the database by the issuance year1 and by the original contractual duration (in months) of the defaulted lease contracts. It follows from it that most of the contracts have a contractual length of four to six years while their allocation throughout years of inception seems to be quite regular with a number of observations that allows us to segment the data further. Table 2 presents a stratication that makes now emphasis on the eective maturity, i.e. focusing on the year of default. Given that the default term will be a key variable for our analysis, it is important to note that years 1990 to 1992 oer little
1

i.e., cohort by cohort.

Credit Risk Mitigation in Auto Leases

Annuity

Interest Time To Maturity

Previous invoice Total overdue Default date Outstanding Capital

Principal

Figure 1: Total overdue and capital outstanding denitions against interest and principal components of the leasing annuity. diversity2 . Would the loss-at-default be inuenced by the age of the contract, an analysis made only on those years will undoubtedly bring the conclusion that earlier years oer only short defaults and bias therefore our study of the behavior of the loss-at-default. One of the key points of the present study is the analysis of the realized payo at default, in terms of exposure-at-default (EAD) as well as the eective age of the contract at that time. In the present database, the default date is recorded when the eective nal default happens. In particular, the nal default may occur later than the contractual maturity in cases where renegotiation took place. This occurs for 217 contracts of our database, as shown in table 3 together with other specicities: recoveries that are higher than the initial investment (14 cases) and default happening on the inception date (1 case only).

2.2

First computations: measuring time-to-default & losses

As mentioned earlier, the default date may be inuenced by the existence of renegotiation schemes. A reasonable alternative to the use of the months-to-default variable would consist in computing the nancial residual duration of the defaulted lease contracts as the fraction of the remaining total amount due over the original investment conceeded by the lessor. Financial duration: F dur = Total amount due . Original investment (2)

It should be noted that the total amount due may be higher than the original investment or cost of the lease. This may happen because of the integratation of
This is not a surprise as we have a database of completed contracts that also started in the time period under scrutiny.
2

Credit Risk Mitigation in Auto Leases

Start year 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 Total (%)

Interval (number of months to maturity) 0-11 12-23 24-35 36-47 48-59 0 1 3 69 196 0 0 10 64 160 4 1 6 81 174 0 1 17 79 149 7 11 13 51 123 0 2 5 51 129 0 23 8 39 87 0 12 12 32 95 0 0 14 59 142 0 4 44 78 204 0 9 37 85 160 0 5 9 42 93 11 69 178 730 1712 0.23% 1.43% 3.69% 15.12% 35.46%

60-71 185 162 165 175 143 147 133 157 213 263 196 126 2065 42.77%

Over 5 0 0 4 10 7 8 6 0 9 12 2 63 1.30%

Total 459 396 431 425 358 341 298 314 428 602 499 277 4,828

(%) 9.51% 8.20% 8.93% 8.80% 7.42% 7.06% 6.17% 6.50% 8.86% 12.47% 10.34% 5.74% 100%

Table 1: Contractual maturity stratication of the data. Number of observations by contract duration (interval in months) and by the year of inception of the lease). unpaid accrued interest fractions of annuities (see gure 1; the dashed zone might be higher than the principal amount depicted by the leftmost triangle). Table 4 presents the overall distribution of our database accross the so-called nancial duration (F dur) and the year of inception. The data appears to be regularly distributed on four equal intervals between 0% and 100% of the original investment. Evidence is provided on 173 contracts that supersedes the 100% value. Losses3 are computed as the total amount due minus the recovered value of the asset being sold at default for contract i, i.e. LGDi = total amount due recovered value. (3)

LGDi takes therefore positive values for losses and negative values for net gains to the lessor. As mentioned earlier, the total amount due does not integrate any other amounts than the unpaid portion of the principal amount of the lease (including the residual value) plus unpaid accrued interest. Moreover, the recovered amount is a pure marked-to-market value as it comprises only the realized value of the leased asset sold at the nal default date. However, this version of the LGD does not take the opportunity cost of the recovery delay into account. The recovery delay is dened as the number of months between the default date and the recovery date. As it can be seen in gure 3 in the appendix, a majority of the contracts shows a delay of around 2 years when excluding contracts with renegotiation schemes.
3

These are of course "losses-given-default" since the database only comprises defaulted contracts.

Credit Risk Mitigation in Auto Leases

Start year 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 Total

Default year 1990 16 16 1991 99 20 119 1992 155 111 35 301 1993 123 119 111 30 383 1994 54 78 157 165 18 472 1995 12 59 92 95 91 23 372 1996 8 30 86 99 92 10 325 1997 5 35 87 88 79 10 304 1998 1 1 13 34 69 77 75 16 286 1999 1 15 44 52 83 67 37 299 2000 8 13 33 40 98 105 36 333 2001 6 29 64 101 218 115 43 576 Total 459 396 431 425 352 335 280 272 282 360 151 43 3,786

Table 2: Default stratication of the data. Data observations distributed by year of inception (start year) and by year of default.

Case description Super-recoveries: recovery > initial investment Certain renegotiation: months to default > months to maturity Instantaneous default: default date = start date

Number of occurrences 14 217 1

Table 3: Special cases. Number of observations corresponding to various special scenarios. We therefore computed a second version of the LGD, where the total due amount is restated to include this opportunity cost, as total amount due = total amount due (1 + yield%)d/12
LGDi

(4) (5)

= total amount due recovered value,

where d is the recovery delay in months and yield% is the yearly yield information (in percent) of every contract used as the opportunity cost base. We decided therefore to undertake the study with both calculations since it would allow us to show the impact of the recovery delay on our estimates4 . The recovery delay may account for the liquidity risk and other operational issues to which the asset liquidation procedure is subject. Table 5 shows that, while the majority of defaulted contracts experience a loss, more than 25% of them generate a net gain for the lessor upon the sale of the asset.

All future references to LGD cover systematically both denitions.

Credit Risk Mitigation in Auto Leases

Start year 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 Total

Interval (F dur) 0-25% 25%-50% 70 107 62 88 60 97 53 87 83 70 67 77 75 78 80 83 110 122 96 119 40 95 13 15 809 1038

50%-75% 156 120 126 119 111 114 79 93 116 251 202 114 1601

75%-100% 109 107 120 132 86 75 62 53 72 130 142 119 1207

Over 17 19 28 34 8 8 4 5 8 6 20 16 173

Total 459 396 431 425 358 341 298 314 428 602 499 277 4828

Table 4: Eective nancial duration. Fraction of the accomplished nancial duration up to the default time, by year of inception. Financial duration = (total amount due/original investment). 2.2.1 LGD relative to the due amount

We also need the expression of the average LGD relative to the total amount due for a given portfolio. This is5 XN XN

LGD%pf

XN

i=1

LGDi

=
1 N

i=1

total amount duei

average portfolio loss , average portfolio due value

1 N XN

i=1

LGDi

(6)

i=1

total amount duei (7)

and is valid for both versions of LGD provided that the total amount due is consistent with the LGD version chosen: LGD% = f (LGD , total amount due ) pf (8)

3
3.1

The Methodology
Loss given default

Starting from the overall database and after verifying for the consistency of the records considered, various sampling criteria were retained according to the empirical
% Even though this expression is straightforward, computing instead the average of LGDi per contract i would result in values and a meaning that may be far from our objective. Since the leasing company will focus on overall losses on a given portfolio rather that its ability to have the lowest loss per contract, we will make use of equation 6 which is equivalent to compute a weighted average % of the loss rate per contract LGDi . 5

Credit Risk Mitigation in Auto Leases

Sign of LGDi >0 =0 <0 LGDi >0 =0 <0 Total (#) Total (%)

Interval 0%-25% 260 9 540 278 2 529 809 16.76%

25%-50% 536 38 464 602 6 430 1038 21.50%

50%-75% 1255 47 299 1356 5 240 1601 33.16%

75%-100% 1081 49 77 1159 2 46 1207 25.00%

Over 169 0 4 169 0 4 173 3.58%

Total (#) 3301 143 1384 3564 15 1249 4828

Total (%) 68.37% 2.96% 28.67% 73,82% 0,31% 25,87% 100%

Table 5: Loss distribution. Number of positive, null and negative losses. Negative values are equivalent to a net gain for the lessor upon the sale of the asset serving as collateral. evidence of the past section: All contracts vs. contracts characterized by F dur [0%, 40%] or by F dur [60%, [. All contracts vs. contracts defaulted in bad years or in good years6 .Three samples will be considered7 : observations of all years taken together, observations of years 1993 and 2001 (considered as the bad years) and observations from 1997 and 1998 (considered as the good years)8 . Those three samples will be generated on the basis of the year of default9 of each contract. Since we want to focus on just the payo at default and its market risk component10 , we precisely need to estimate the weight of systematic risk in the residual value of the lease11 . Two sets of criteria generate six dierent samples for the study of the empirical LGD%pf distribution.
It must be reminded that table 2 shows that earlier years present a limited data stratication that points them out automatically as having rapidly defaulted. Therefore, data belonging to the period 1990-1992 was dropped out. Any further study sampled on the basis of this ranking would therefore need to consider only years from 1993 onwards. 7 These years or annual periods were determined based on the results presented by Linna [13] and a discussion by Cooper [7] on the paper of Bergman, Bordo and Jonung [3]. An examination of Belgian stock index series was also performed. 8 By choosing only 1998 as the good year, we ensure the compatibility of the samples accross the LGD measures. 9 In opposition to the inception date of the contract. Using the latter would implicitely mean somehow that a contract LGD would be hopelessly predictable based on the state of economy at origination regardless of the time-to-default. Such a world would be scaring. 10 i.e., the asset liquidation at a marked-to-market price. 11 The New Basel Accord requires frequent analysis to be conducted ex-post on past data, the information on default dates is an observable variable.
6

Credit Risk Mitigation in Auto Leases

Additional cases are also considered to study the impact of negative losses (i.e. net gains for the lessor: LGD < 0) and the implications of the renegotiations operated on some contracts (months-to-default > contract duration). A bootstrapping methodology will allow us to recover a consistent distribution of LGD%pf from our empirical samples, by iteratively and randomly choosing portfolios of contracts. The generated distributions should reect the properties of the true distribution of losses.

3.2

Residual value risk

Originally, the residual value risk is the banks exposure to potential loss due to the fair value of the asset declining below its residual estimate at lease inception (see [2]). This residual estimate is contractually dened such that the lessee may keep the asset at the end of the contract in exchange of its payment, which represents the strike price of an option that the lessee may or may not exercise. Given that this residual value is xed ex-ante, there is no uncertainty about its amount. The only remaining uncertainty is the liquidation value of the collateral asset. Additionally, one of the main characteristics of the leasing practices on the national market on which the database is provided is that the option value represents only a tiny fraction of the original investment12 . The diculty in obtaining consistent estimations of collateral market values for non-defaulted contracts is here encompassed by observing eective losses on defaulted contracts that are close to maturity. In this context, we argue that the distribution of LGDs would allow us to infer the residual value risk of lease contracts. This makes the implicit assumption that if the recovered value is enough to cover the small fraction of total amount still due, then we are left with the rest to cover the residual value. That is why we will focus on defaulted contracts that are close to their contractual maturity. The overall sample of LGDs will act as a proxy for the residual value risk for (a) contracts that are completely honoured but where the option is not exercised by the lessee and (b) for defaulted contracts13 .

3.3

A bootstrapping framework

In the present sample, the loss rate distribution of a sub-portfolio is estimated by a resampling method called bootstrap". Bootstrap is a powerful econometric technique that allows to approach the true distribution by making equiprobable draws with replacement from an observed sample of data. The advantage of this method is that it is non-parametric and relies only on observed data. We will perform here the same application of this technique as in Carey [5]. The basic process consists of
An examination of an overall database of 57265 completed contracts from the same source (defaulted and non-defaulted) shows that the option or residual value accounts for 3.1% of the original investment on average with a standard deviation of 1.87%. 13 We forego the cases where the lessee does exercise her option but this does not break our starting motivation as it makes us priori even more conservative looking at residual value as a potential loss-risk.
12

Credit Risk Mitigation in Auto Leases

10

choosing randomly, with replacement, a portfolio of n lease contracts for a randomly chosen year. The draw of a year can be interpreted as a draw from the best available representation of the possible macroeconomic conditions inuencing the risk factor. The assumption is that each year has the same probability of being drawn (e.g. if we have 6 observation years, each year has a 1/6 probability of being drawn). The process is iterated i times. When a lease is drawn, then the associated loss is the dierence between the due amount and the recovered value. LGD can be either negative or positive. In the latter case, the recovery rate is higher than 100%. A single iteration of the procedure yields a loss rate for a given state of the economy (or a given year). Using a large number of iterations enables us to obtain a probability distribution of loss rates as a percentage of the total amount due. In our case, we ran the simulation procedure for portfolios consisting of 100, 400, 1600 and 6400 lease contracts by carrying out 100,000 iterations. By performing the draw procedure in two stages (i.e. drawing rst a year and then a portfolio of n leases), we avoid the understating of tail loss rates. Otherwise, the combination of default experiences from dierent years would lead to a tricky mixture of the underlying systematic factors and hence to over-diversication14 . We assume that the portfolio has a high degree of granularity, so that no single lease represents more than a small fraction of our portfolio15 .

4
4.1

The Results
Loss given default

Losses-given-default (LGD%pf ) have been tested against several factors: the belonging to any of the bad or good years (as dened in subsection 3.1), portfolio diversication (represented by the portfolio size variable), and the importance of the payments not honoured compared to the initial invested amount (F dur). Table 6 shows that, median values for good years are better than those for the bad years. But, dierences in this direction vanish when considering higher percentiles. Moreover, there seems to be an anticyclality eect in the extreme values of the distribution. That might be explained by the the fact that more people turn towards the automotive secondary market in bad business cycles and/or that the composition of the automotive portfolio changes16 . But the reasons linked to the market of the asset collateral will be claried when considering samples with contracts that are close to maturity since, then, we will be facing LGDs whose bigger component is the recovery part (see the results in table 7). This is indeed part of our motivation in examining the results based on residual values since we attempt there to concentrate on the market risk component (see subsection 4.2).
Indeed, when taking contrats issued on dierent years in the same portfolio, business cycle eects could compensate across years. 15 The algorithm is written in MATLAB and can be shown upon request. 16 Some categories of automotive assets are less liquid than others. If they are more numerous in good years, that could explain the eect in the right hand tail of the distribution.
14

Credit Risk Mitigation in Auto Leases

11

We need to mitigate somehow our comment on median values by the fact that standard deviations seem to shed light on a lower diversication in the original data for smaller samples17 . However, as far as we are concerned with conservative requirements and therefore high percentiles, the extreme values of the bad years distribution do not show any loss far beyond those of the overall sample nor for the good subsample. The higher kurtosis for the bad years sample express the fact that there is more power in the tails of the distribution for that subsample which conrms the need to look at high percentiles for capital requirements. Also, the positive skewness of the overall sample contradicts the negative one of the subsamples which advocates for a higher potential of gains in the subsamples.
Case & Sample criteria Portfolio size 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 Median 16.00% 15.94% 16.14% 16.64% 17.80% 17.99% 18.14% 18.12% 16.26% 16.24% 16.21% 16.17% 20.74% 20.87% 20.99% 21.01% 22.95% 22.63% 22.34% 22.16% 18.92% 18.91% 18.88% 18.82% 95% 26.60% 26.60% 27.33% 29.33% 27.12% 27.06% 27.21% 28.63% 38.52% 38.78% 39.29% 40.24% 28.33% 28.42% 29.44% 32.46% 29.83% 30.02% 30.61% 32.60% 40.74% 41.00% 41.49% 42.44% Percentiles 99% 27.28% 27.84% 29.45% 33.28% 27.84% 28.50% 29.91% 32.81% 39.06% 39.88% 41.46% 44.58% 29.01% 29.63% 31.90% 37.41% 30.47% 31.12% 32.69% 36.14% 41.28% 42.08% 43.64% 46.68% 99.9% 27.80% 28.83% 31.58% 37.13% 28.42% 29.65% 32.16% 37.23% 39.58% 40.84% 43.60% 48.60% 29.53% 30.68% 36.24% 48.40% 31.00% 32.10% 34.53% 39.85% 41.78% 43.02% 45.74% 50.60% Standard deviation 5.10% 5.22% 5.67% 7.25% 16.13% 16.19% 16.46% 17.59% 17.65% 17.70% 17.91% 18.70% 3.72% 3.94% 4.70% 6.83% 16.44% 16.50% 16.76% 17.84% 17.11% 17.17% 17.38% 18.20% Skewness 0.52 0.49 0.37 0.01 -1.92 -1.91 -1.90 -1.89 -0.14 -0.15 -0.19 -0.35 0.15 0.16 0.20 0.13 -1.89 -1.89 -1.87 -1.87 -0.13 -0.14 -0.19 -0.35 Kurtosis 2.01 2.10 2.37 3.02 5.69 5.71 5.76 6.34 1.72 1.75 1.85 2.27 2.67 2.58 2.72 3.55 5.51 5.53 5.58 6.13 1.74 1.76 1.87 2.29

LGD%pf
1 All years (4391 obs.) 2 Years 1993, 2001 (959 obs.) Years 1997, 1998 (583 obs.)

LGD% pf
1 All years (4391 obs.) 2 Years 1993, 2001 (959 obs.) 3 Years 1997, 1998 (583 obs.)

Table 6: LGD%pf results by LGD cycles. Bootstrap statistics for three samples based on dierent years given their ranking in terms of best to worst average LGD (not in terms of probability of default like other studies). The best year is 1998 while the worst ones are 1993 and 2001. These two samples are compared to the overall one to check for cycle eects. The size of the portfolio is the number of contracts bootstrapped for each bootstrapped year. Focusing now on the nancial remaining duration and according to the results shown in table 7, the more contracts the portfolio contains, the less LGD distributions deviate from the mean loss. Moreover, in the case of old contracts (F dur < 40%)18 , the 99.9 percentile is negative for portfolios of 1600 and 6400 contracts while it is pos17 18

This is maybe why the overall sample shows a better median than any of the subsamples. i.e. contracts with a long track record.

Credit Risk Mitigation in Auto Leases

12

itive for smaller portfolios. The portfolio diversication eect is thus well established for credit risk in the automotive leasing business. This conrms the importance of the granularity and the low value of individual exposures criteria retained in the regulatory capital framework of the New Basel Capital Accord (see [2], 44).
Case & Sample criteria Portfolio size 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 Median 17.37% 17.53% 17.59% 17.83% -30.90% -30.99% -30.86% -30.42% 26.77% 26.76% 26.79% 26.87% 22.59% 22.71% 22.72% 22.54% -23.52% -23.82% -24.49% -25.11% 31.31% 31.45% 31.59% 31.30% 95% 26.60% 26.62% 27.34% 29.36% -11.65% -10.95% -8.96% -4.02% 34.27% 34.22% 34.46% 35.64% 28.37% 28.65% 29.89% 33.12% -7.29% -6.53% -4.59% -0.03% 38.42% 38.10% 38.29% 40.31% Percentiles 99% 27.27% 27.85% 29.47% 33.25% -9.83% -7.81% -3.48% 4.81% 34.80% 35.25% 36.19% 38.54% 29.01% 29.70% 32.07% 37.80% -6.16% -4.40% -0.35% 7.58% 39.95% 41.20% 43.60% 48.34% 99.9% 27.78% 28.88% 31.34% 37.63% -8.10% -4.52% 2.33% 14.12% 35.22% 36.11% 37.78% 41.55% 29.51% 30.80% 36.67% 48.53% -5.18% -1.92% 4.75% 16.17% 41.23% 43.49% 48.46% 57.17% Standard deviation 4.45% 4.57% 5.02% 6.52% 60.00% 60.20% 60.99% 63.68% 4.16% 4.23% 4.47% 5.35% 3.99% 4.17% 4.85% 6.76% 59.26% 59.45% 60.18% 62.69% 4.52% 4.61% 4.93% 6.01% Skewness 0.43 0.42 0.35 0.08 -2.25 -2.26 -2.29 -2.41 -0.10 -0.10 -0.10 -0.28 -0.42 -0.35 -0.18 0.00 -2.29 -2.30 -2.33 -2.43 -0.34 -0.28 -0.09 0.11 Kurtosis 2.13 2.22 2.46 3.07 6.89 6.97 7.30 8.54 2.80 2.80 2.84 3.80 2.46 2.39 2.53 3.43 7.07 7.15 7.45 8.61 2.54 2.68 3.21 4.80

LGD%pf
1 All sample (4827 obs.) 2 0<Fdur<40% (1378 obs.) 3 Fdur>60% (2353 obs.)

LGD% pf
1 All sample (4827 obs.) 2 0<Fdur<40% (1378 obs.) 3 Fdur>60% (2353 obs.)

Table 7: LGD results by F dur. Bootstrap statistics for three samples based on dierent intervals of nancial duration. The nancial duration is computed as the ratio of the total due at default over the initial investment. The size of the portfolio is the number of contracts bootstrapped for each bootstrapped year. This table is based on the original sample of 4,828 observations. Table 7 also presents important results concerning the F dur variable: 99,9% of the generated portfolios present a maximum LGD higher in the case where F dur is more than 60% than in the general case (35.22% vs. 27.78%, for LGD%pf ). More importantly, LGDs are negative for well diversied portfolios including contracts whose F dur is less than 40%. That makes the maximum LGD of 99,9% of these portfolios much smaller than that of 99,9% of all portfolios taken together. This evidence can be explained by the linear timely reduction in the total amount due and the recovered value crossing the latter at 2/3 of the contract completion as illustrated in gure 2. Therefore, contracts with a low remaining duration are undoubtedly contracts that present a much smaller credit risk component, getting closer to their residual value which is contractually xed. Would the lessees default on this value or neither default nor exercise their option, gure 2.shows that the problem has shifted to one of market risk. This provides a visual motivation to study

Credit Risk Mitigation in Auto Leases

13

more precisely LGDs on old contracts (in F dur terms) as proxies for net residual values.

Figure 2: The gure plots the average total due and the average recovered value as fractions of the original investment, against the percent of nancial completion of the contract (F dur) dened as the ratio of the total due over the investment level. The same conclusions apply to LGD% . Indeed, relationships between subsampf ples are exacerbated. For the time being, any discrimination by a F dur interval has however not yet been considered in the regulatory capital requirements framework. 4.1.1 Special cases

In this section, we will focus on the inuence of negative LGD of several contracts (i.e. gains resulted from a higher recovered value than the total due amount) and the impact of potential renegotiations (contracts whose contract duration is shorter than months-to-default period can be considered as renegociated contracts) on LGD distributions. By withdrawing these special contracts from the samples, the newly obtained LGD distributions emphasize the huge benecial inuence of positive LGD and the small positive impact of renegotiation capacity. Indeed, 50% of the portfolios with positive LGD present a maximum LGD

Credit Risk Mitigation in Auto Leases

14

(29,55%) much higher than the maximum LGD (17,37%) of half all contracts together. Secondly, the maximum LGD of half all contracts together is slightly smaller (17,37%) than the maximum LGD of 50% of non renegotiable contracts(18,13%).

Case and sample criteria LGD > 0 Months-to-default < Contract duration

Median 29.55% 18.13%

95% 37.66% 26.60%

99% 38.34% 27.29%

99.9% 38.87% 27.80%

Standard deviation 3.64% 4.10%

Skewness 0.59 0.43

Kurtosis 2.02 2.23

Table 8: Bootstrap statistics for two particular samples: a sample that does not integrate the cases where the LGD is positive and a sample free from renegotiations. Renegotiations plans are responsible for the existence of contracts where the nal default occurs after the original contractual maturity. Both results were obtained from bootstrapping yearly portfolios of 6400 contracts.

4.2

Residual value losses

We approximated residual value risk by using a sample of all contracts whose F dur is less than 20% and that default, i.e. we forego contracts where the option is exercised at the end of the contract. Therefore, we overestimate the resulting average eective residual value (conservative behavior). Table 9 presents values consistent with previous results on the sampling based on a low F dur (see table 7). Indeed, residual values are completely characterized by net gains, which means that we have demonstrated that residual value risk is inexistent in our database irrespectively of the year of default (at all percentile levels). Thus, taking into account exercised-option cases would lead us to the same statement as they have a null residual value. Moreover, and consistently with our previous comment on the specic behavior of low nancial duration contracts., we obtain now a clear cyclical eect in portfolios made of those contract. Indeed, the 99.9th percentile for the bad years sample shows a lower gain than that for the good years sample. Typically, for contracts where the total amount due is close to the residual option value of the lease, the asset collateral represents almost a pure market risk exposure and, depending on the state of the secondary automotive market, may generate a strong positive earning to the lessor. Results on the LGD% panel present the same pattern. pf Thus, focusing on short time-to-maturities (in nancial terms), we shifted from a small anticyclical evidence on losses to a cyclical evidence on gains. But the latter are based now on LGDs whose total amount due component represents only a tiny residual credit risk exposure, the market value of the asset collateral more than compensating it in all cases. Therefore, while any drawn conclusion on the role of business cycles would be dicult to objectively implement in a regulatory framework19 , F dur stands as an
19

It depends on the criterium used to dene a given year as bad or good. Moreover, for pruden-

Credit Risk Mitigation in Auto Leases

15

objective driver that intrinsically represents the engineering framework of the lease industry: combining contracts with diverging amortization schemes to increase diversication and to enable a continuous revolving of lent amounts.

Case & Sample criteria

Portfolio size 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100 6400 1600 400 100

Median -1.53 -1.52 -1.52 -1.52 -2.06 -2.17 -2.42 -3.03 -2.82 -2.98 -3.01 -3.02 -1.41 -1.41 -1.43 -1.45 -2.00 -2.11 -2.35 -2.94 -2.52 -2.52 -2.53 -2.57

95% -0.71 -0.71 -0.69 -0.63 -0.33 -0.33 -0.33 -0.28 -1.33 -1.31 -1.27 -1.20 -0.66 -0.65 -0.63 -0.56 -0.32 -0.32 -0.31 -0.26 -1.29 -1.28 -1.24 -1.17

Percentiles 99% 99.9% -0.70 -0.67 -0.62 -0.50 -0.31 -0.29 -0.24 -0.14 -1.29 -1.24 -1.15 -0.97 -0.64 -0.62 -0.57 -0.44 -0.29 -0.27 -0.22 -0.12 -1.26 -1.21 -1.12 -0.94 -0.68 -0.64 -0.55 -0.35 -0.29 -0.25 -0.17 -0.01 -1.26 -1.18 -1.03 -0.78 -0.62 -0.58 -0.49 -0.28 -0.28 -0.23 -0.16 0.01 -1.23 -1.15 -1.01 -0.76

Standard deviation 1.93 1.94 1.94 1.98 7.13 7.13 7.13 7.14 6.49 6.50 6.50 6.51 1.84 1.84 1.85 1.88 7.07 7.07 7.08 7.08 6.41 6.41 6.41 6.42

Skewness -2.35 -2.35 -2.34 -2.31 -1.74 -1.74 -1.74 -1.73 -1.69 -1.68 -1.68 -1.66 -2.32 -2.32 -2.31 -2.27 -1.82 -1.82 -1.82 -1.81 -1.70 -1.70 -1.69 -1.67

Kurtosis 7.64 7.65 7.66 7.70 5.02 5.02 5.01 4.98 4.02 4.01 4.00 3.96 7.52 7.53 7.53 7.52 5.32 5.32 5.32 5.29 4.03 4.03 4.02 3.98

LGD%pf
1 All years (628 obs.) 2 Years 1993, 2001 (91 obs.) 3 Years 1997, 1998 (86 obs.)

LGD% pf
1 All years (628 obs.) 2 Years 1993, 2001 (91 obs.) 3 Years 1997, 1998 (86 obs.)

Table 9: Residual values distributions. Bootstrap statistics on residual values estimated with contracts which Fdur is lower than 20number of contracts bootstrapped for each bootstrapped year. Observations of obtained samples are provided in parentheses.

Conclusion

First, the present study proposes a systematic bootstrap methodology which, in the case of a retail automotive leasing portfolio, has allowed us to verify the relevance of some key drivers in the analysis of the LGD. The stratication of the data in terms of aging of the contracts is central to the obtained distribution of LGD. The present study has shown that comparing samples without taking this into consideration could substantially bias our conclusion. Moreover, the nature of credit risk and the ability of the physical collateral to substitute a market risk to the original credit risk exposure is strongly dependent on the remaining nancial duration of the contract.
tial purposes, regulatory authorities are always reluctant to free up limits based on the anticipation of better years.

Credit Risk Mitigation in Auto Leases

16

Indeed, that is the reason that led us to use those LGDs as proxies for the residual value risk. Therefore, such a result would advocate for a thinner segmentation of the risk weighting than as it is proposed under the New Basle Accord [2]. The current consultative document proposes a unique weighting of 75% in the standardized approach if sucient conditions are met (see [2], 43 & 44). Otherwise, the ratio is 100%. This methodology may help managers in switching to a more advanced approach and to screen their activity log to extract its stratication map. Secondly, studying LGDs has led us to conclude that we can anticipate substantially lower losses towards the end of the contract, as measured in fraction of the original investment still current (F dur). The advantage of this measure as opposed to one based on calendar date dierences between the inception date and the default date, is that it is a pure nancial measure that is independent from the contractual date specicities of each sample. Moreover, using F dur provides a natural cash-ow weighting of the times-to-default. Third, using only contracts that are close to their completion, and based on the previous results, we drew conclusions on the residual value risk. Results show that this exposure net of the recovered value is nally protable for the lessor, which would mean that the lessor is better o in cases where the lessee does not exercise her option. Again, this is inherent to the fact that old contracts are contracts with low exposures left. This advocates for a treatment of the problem in one single block, without a separation of the credit and market risk exposures since there are intrinsically linked. Extensions to the present research would include: The integration of other types of collateral assets (as equipment) that show a much wider diversity. As cited in [17], Rolling stocks register the highest recovery rates because leasing specialists benet from a good understanding of secondary markets as they are well organized and homogeneous. Studying more consistently the explanatory power of each key variable on LGDs. Finally, the present study contributes to the enhancement of the public awareness on retail nancing risk and the eective mitigation provided by physical collateral thanks to a proprietary database. Given the results, showing evidence on these data could only undoubtedly benet leasing companies as a whole when discussing with the Basle regulatory body.

References
[1] Allen L., DeLong G. and Saunders A. (2003), Issues in the Credit Risk Modeling of Retail Markets, paper presented at the Conference on Retail Credit Risk Management and Measurement, Federal Reserve Bank of Philadelphia, April 24-26, 2003.

Credit Risk Mitigation in Auto Leases

17

[2] Basle Committee on Banking Supervision (2003), The New Basle Capital Accord, Third Consultative Document, issued for comment by July 2003, 226 pages. [3] Bergman M., M. Bordo & L. Jonung (1998), Historical Evidence on Business Cycles: The International Evidence in Fuhrer, Jerey C. and Scott Schuh (eds.) Beyond Shocks: What Causes Business Cycles?, Conference Series No. 42, pages 65-113, Federal Reserve Bank of Boston. [4] Calem, P., LaCour-Little, M. (2003), Risk-based capital requirements for mortgage loans, Journal of Banking and Finance. forthcoming. [5] Carey M. (1998), Credit Risk in Private Debt Portfolios, The Journal of Finance, Vol. 53, No. 4, Papers and Proceedings of the Fifty-Eighth Annual Meeting of the American Finance Association, Chicago, Illinois, January 3-5, 1998. (Aug., 1998), pp. 1363-1387. [6] Carey, M. (2000), Dimensions of credit risk and their relationship to economic capital requirements, Working Paper 7629. National Bureau of Economic Research, March. [7] Cooper, Richard (1998), Discussion, Harvard Business School, Discussion paper presented at a conference of the Federal Reserve Bank of Boston in June 1998: Beyond Shocks: What Causes Business Cycles?, 6 pages. [8] Credit Suisse Financial Product (1997), CreditRisk+TM: A Credit Risk Management Framework, Credit Suisse. [9] De Laurentis, G., Geranio, M. (2001), Leasing recovery rates, LeaseuropeBocconi University Business School Research Paper. [10] Franks, Julain R. and Stewart D. Hodges (1978), Valuation of Financial Lease Contracts: A Note", The Journal of Finance, vol.33, May 1978, pp. 657-669. [11] Lease, Ronald, John J. McConnell and James S. Schallheim (1990), Realized Returns and the Default and Prepayment Experience of Financial Leasing Contracts", Financial Management, Summer 1990, pps 11-20. [12] Leaseurope, (2002), Leasing Activity in Europe. Key Facts and Figures, http://www.leaseurope.org/pages/Studies_and_Statements/. [13] Linaa, Jesper Gregers (2003), On the Conformity of Business Cycles Across EU Countries, Working paper, University of Copenhagen and EPRU (Economic Policy Research Unit), 34 pages. [14] Maddison Angus, The World Economy: A Millennial Perspective. Paris: OECD, 2001. 384 pages. [15] McConnell, John J.and James S. Schallheim (1983), "Valuation of Asset Leasing Contracts", Journal of Financial Economics, 12, 1983, pp. 237-261.

Credit Risk Mitigation in Auto Leases

18

[16] Pykhtin M.and Ashish Dev (2003), Residual risk in auto leases, Risk Magazine, October 2003, 5 pages. [17] Euronas and Leaseurope Joint Annual Conference, Salzburg, 2002, September 22nd-24th. [18] Schmit, M., (2003), Is automotive leasing a risky business?, Working Paper 03/009, Centre Emile Bernheim, Solvay Business School, Universit Libre de Bruxelles. [19] Schmit, M. (forthcoming), Credit risk in the leasing industry, forthcoming in the Journal of Banking and Finance, 24 pages. [20] Schmit, M. and J. Stuyck (2002), Recovery rates in the leasing industry, Working Paper presented at Leaseuropes Annual Working Meeting, Salzburg.

Figure 3: Recovery delay distribution. The gure presents the frequency distribution of the delay in the database, i.e. the number of months between the default date and the recovery date.